Abstract
To create good AI, we must simplify it, and to simplify it, we must improve our techniques. In my capstone research, I aim to increase the effectiveness and quality of AI-adjacent methods in tasks where AI has most recently been found to achieve the state-of-the-art. To do this, I identify theoretically-grounded strategies for simplifying how AI models make decisions. In my STS research, I am to better understand the broader context of AI’s development. I map out the broad range of interpretations people have had for AI models, what tensions there could be when groups have different priorities for what the order of priority should be for AI’s development, and what design principles can improve the design of AI systems. These topics are highly connected because creating AI that is more easily understood provides a strategy for improving its capabilities and aligning it with desirable principles. I suggest that rather than treating AI alignment and capabilities research as separate, that one must get a better picture of what capabilities exist or can be created, in case they provide new opportunities for alignment that would not exist from a more abstract perspective alone.
My capstone project centers around designing better models for teams of robots trying to coordinate with each other to navigate to goal locations without colliding. This problem space is broadly referred as multi-agent pathfinding, and currently used in scenarios with collections of robots on warehouses. Eventually, if humans and robots (of any type, even Roombas) occupy the same spaces, technologies like this will allow robots to efficiently plan how to navigate a scene without getting in the way of humans. Traditional approaches tend to require simplified assumptions, and more recent approaches have turned to machine learning models, and particularly diffusion models, to push the frontier of performance.
My research finds that in some cases, one can improve the efficiency of diffusion models by exploiting one’s knowledge of how the task is structured. In fact, one can improve performance to such an extent that the information a diffusion model would typically have provided, after training for several hours on a set of demonstrations, can be computed directly and mathematically, while still outperforming classical approaches. My theoretical formulation applies to a general characterization of tasks that extends beyond the multi-agent pathfinding regime. Mathematically, it exploits the independence structures of Markov decision processes, which have probabilistic graphical models that form a straight line (i.e., what to do at a given point in time depends purely on the state of the system at the previous point in time). Therefore, it has the potential to improve the performance of AI models in a wide range of domains, which I hope to continue as future work this summer.
In my STS paper, I broadly investigate the different interpretations of how AI should be used, how aligned these interpretations are with real user well-being, and what actors may influence the dynamics of stabilization. AI is clearly ubiquitous, and therefore it is important to understand how the role AI plays in our lives gets decided. I study this using the Social Construction of Technology (SCOT) framework, which identifies different interpretations of a technology and what social groups push for different types of adoption. I also look at what has already happened with AI, to see if AI is being designed in a manner consistent with users’ best interests.
Development is primarily driven by private labs and government entities, potentially with a disregard for how the models impact consumers. Concerningly, I find that the Department of War advocates for deprioritizing AI alignment concerns for the purpose of faster development. Simultaneously, I found several case studies indicating the adverse effects when an AI model is created without proper underlying principles, including examples where young users have committed suicide in periods of social isolation concurrent with chatbot use, and therapists find an increase in patients who claim to have made grand scientific discoveries because a chatbot supported their beliefs. I conclude that the design of AI models may have underacknowledged second-order effects, and that therefore engineers who create AI models have a uniquely high-leverage position over their users. Engineers should take time to understand the possible implications of their modeling choices, rather than treating adverse outcomes as acceptable risks. Rather than this needing to be a problem, I present qualitative results suggesting that fundamental research and alignment research can be complimentary, through an example of a “moderator” chatbot that detects red flags and intervenes in conversation.
My hope is that my current capstone research provides some insights into how we can create more interpretable AI models, by investigating classical approaches more closely and deriving insights for creating models that do not need as much data to reach a desired performance level.
